Prediction of Soil Total Nitrogen Content Using Spectraradiometer and GIS in southern Iraq
Abstract
In this study, soil samples were collected from two locations: Samawa and Rumetha in southern Iraq. The samples from each location were split into two datasets: calibration set and validation set. VNIR reflectance (350-2500 nm) and GIS-Kriging were used in combination with Partial Least Square (PLS) to predict total N. only two regions reported higher determination coefficient R2 and lower Root Mean Square Error (RMSE) than the other wavelength regions. PLS calibration models yielded an R2 of 0.96 and 0.97 for Rumetha and 0.87 and 0.94 for Samawa location in bands at 500-600 and 800-1000 nm, respectively. The potential of VNIR-based and GIS-Kriging models to predict new unknown soil samples were assessed by using validation datasets from both studied locations. The cross-validation of GIS-Kriging models were unsatisfactory predicted with an Q2 of 0.28 between laboratory-measured and predicted total N values for Rumetha and 0.43 for Samawa location. While VNIR- based validation models achieved highly predictive power with an R2 v of 0.84 between laboratory-measured and predicted total N values for Rumetha and 0.85 for Samawa location. These results reveal extremely decreasing in model predictive ability when shifting from VNIR Spectroscopy method to GIS-Kriging.
Keywords
Download Options
Introduction
Visible Near Infrared Radiation (VNIR) is a tool for molecular structure determination (Pasquini, 2003). It's concept based on the interactions between the investigated sample and the VNIR at the region of electromagnetic spectrum in range of 400- 2500 nm (Canasveras et. al., 2012).
There are several factors contribute in the bandwidth of infrared absorptions (narrow or broadening). Collision between molecules and the limited of timelife of an excited state are the main sources of line broadening, where the less well describe energy associated with the shorter lifetime of transition to the excited state (Stuart, 2004). In spite of that the absorption intensity is highly related to the change in molecular dipoles, hence, the large change lead to a very strong absorption and inversely, a very weak absorption associated with a very small change in dipole (Günzler and Gremlich, 2002; Hollas, 1996).
Over the last 30 years there has been a growing interest in VNIR-Spectroscopy as a potential analytical technique for studying materials in many different fields such as Agriculture, food, textiles, polymers, wool, biomedical and pharmaceutical (Guerrero, et al., 2010; Niemoller and Behmer, 2008). VNIR-Spectroscopy has attracted much attention from soil scientist as a promising technique to determine a lot of soil properties from a single scan, reduce sample preparation, diminish the hazardous of using chemicals, and measurement can be taken both in laboratory and field with a few seconds (Ben-Dor and Banin, 1995; Chang et al., 2001; Viscarra Rossel et al., 2006). Soil spectrum is a characteristic shape that is caused by absorbing light to various degrees depending on the constituents of soil. Absorption of energy occurs mainly due to electrons transition between molecular orbits in visible region and vibrations in molecular bonds in near infrared region (Wetterlind et al., 2013). The frequencies at which light are absorbed match the difference between two energy levels and are displayed in %R (or transformed to absorbance) for analytical purposes (Miller, 2001). Much research has been widely using the diffuse reflectance ( VNIR-Spectroscopy) to determine soil texture, structure, soil organic matter (SOM), fertility, clay mineralogy, and microbial activity (Bowers and Hanks, 1965; Brown et al., 2006; Stenberg et al., 1995).
The soil spectra is characterized by few, broad, and overlapping absorption features due to several reasons such as sample chemical constituent, neighboring functional groups, hydrogen bonds (Miller, 2001). Diffuse reflectance of the soil is mostly influenced by Soil particle size, surface structure, and water films on soil surface (Twomey et al., 1986).
The assessment of soil fertility is usually routine work associated with soil nutrients level and organic matter content as main soil properties for either to Precision agriculture or to maintain soil from degradation (Gobeille et al., 2006). Such estimation needs to acquire data, since the soil sampling and laboratory analysis are costly and time consuming. At this point, spectroscopic techniques become a promising method to make rapid laboratory soil analysis as well as on-line field soil analysis with aid of a field portable instrument (Chang et al. 2001; Pirie et al. 2005; Brown et al. 2006; Nanni and Dematte, 2006). As described by Desbiez et al. (2004), soil fertility is function of soil properties therefore; it is comprehensive concept more than to measure directly.
Soil nitrogen and organic carbon have generally yielded most optimistic correlation coefficients (R2 higher than 0.90) between the actual and estimated concentrations. Dalal and Henry (1986) studied the prediction of total soil nitrogen based upon selection of the combination of three wavelengths 1702, 1870 and 2052 nm. The correlation coefficient of prediction was R2 = 0.92 and the standard errors for the prediction were much larger for soil samples with low contents of organic matter and total nitrogen. Wavelengths range1100- 2300 nm reported to be the best wavelengths to predict soil total N with correlation coefficient R2 = 0.94 (Reeves and McCarty, 2001).
The sensitivity of N and C to infrared radiation is the main driving force that pushes researchers toward soil N and organic C analyses, which proved in very successful calibration coefficient R2 in the region of 0.80 - 0.98 (Reeves et al., 2001; Cozzolino and Moron, 2006; Stevens et al., 2008). Several parameters have been evaluated in addition to total N and C, microbial C and N with reported R2 in the range 0.60 to over 0.90 (Change et al., 2001; Ludwig et al., 2002). In regarding of C and N mineralization: the accumulated mineral N under aerobic and anaerobic incubation conditions have been studied with promising R2 values between 0.70 and 0.80 (Palmborg and Nordgren, 1993). Others studies have reported less correlation values (R2 ˂ 0.5). The disappointing results have been attributed to the small subset samples, diverse locations and/or predicting one experiment with a calibration on the other (Reeves et al., 1999; Change et al., 2005). TerhoevebUrselmans et al., (2006) underlined the importance of the sufficient sample sets to predict biological properties with NIR spectroscopy. The less reliable prediction with very low correlations (R2 <0.50) is usually related to the low concentration or ephemeral and temporal changes in soil solution chemistry (Janik, et al., 1998).
Several factors influencing soil N content, soil particle-size fractions is one of these factors which has been studied by an indirect measurement of soil available N by using NIR spectroscopy (Barthès et al., 2008). The infrared absorption bands are often overlapped, some of interference is associated with soil carbonate contents which is considered as another factor influencing the accuracy of soil N content (Linker et al., 2005; Jahn et al., 2006). Borenstein et al. (2006) considered carbonate overlapping as the largest errors for nitrate determination in calcareous soils, where nitrate band is disturbed by absorbance band of carbonate
Conclusion
In aid of qualitative interpretation (PLS), the spectral bands that showed good correlations with total N have been identified. For total N, only two regions reported high R2 and low RMSE: for bands at 500-600 nm with R2 =0.95 for Rumetha and 0.87 for Samawa and 800-1000 nm with R2 = 0.97 for Rumetha and 0.94 for Samawa location,
The reported wavelength regions was well-known as N. At all reported bands,
The performance of both prediction model: Kriging- based models and VNIR-based models have been assessed by using validation set. In terms of all statistically evaluation parameters, prediction capability of the GIS- Kriging models for total N was as much lower and poor in comparison with VNIR-based model.
